Re: [PATCH net-next] net: dsa: User per-cpu 64-bit statistics

From: Eric Dumazet
Date: Fri Aug 04 2017 - 13:12:01 EST


On Fri, 2017-08-04 at 08:51 -0700, Florian Fainelli wrote:
> On 08/03/2017 10:36 PM, Eric Dumazet wrote:
> > On Thu, 2017-08-03 at 21:33 -0700, Florian Fainelli wrote:
> >> During testing with a background iperf pushing 1Gbit/sec worth of
> >> traffic and having both ifconfig and ethtool collect statistics, we
> >> could see quite frequent deadlocks. Convert the often accessed DSA slave
> >> network devices statistics to per-cpu 64-bit statistics to remove these
> >> deadlocks and provide fast efficient statistics updates.
> >>
> >
> > This seems to be a bug fix, it would be nice to get a proper tag like :
> >
> > Fixes: f613ed665bb3 ("net: dsa: Add support for 64-bit statistics")
>
> Right, should have been added, thanks!
>
> >
> > Problem here is that if multiple cpus can call dsa_switch_rcv() at the
> > same time, then u64_stats_update_begin() contract is not respected.
>
> This is really where I struggled understanding what is wrong in the
> non-per CPU version, my understanding is that we have:
>
> - writers for xmit executes in process context
> - writers for receive executes from NAPI (from the DSA's master network
> device through it's own NAPI doing netif_receive_skb -> netdev_uses_dsa
> -> netif_receive_skb)
>
> readers should all execute in process context. The test scenario that
> led to a deadlock involved running iperf in the background, having a
> while loop with both ifconfig and ethtool reading stats, and somehow
> when iperf exited, either reader would just be locked. So I guess this
> leaves us with the two writers not being mutually excluded then, right?

You could add a debug version of u64_stats_update_begin()

doing

int ret = atomic_inc((atomic_t *)syncp);

BUG_ON(ret & 1);


And u64_stats_update_end()

int ret = atomic_inc((atomic_t *)syncp);

BUG_ON(!(ret & 1));


We probably could have a CONFIG_DEBUG_U64_STATS that could be used on
64bit kernels as well...